Does the thundering herd issue still exist in epoll of new Linux kernel?

Question

I saw that fs/eventpoll.c in the kernel source code is written like this:

static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
           int maxevents, long timeout)
{
....
        init_waitqueue_entry(&wait, current);
        __add_wait_queue_exclusive(&ep->wq, &wait);    // *** NB
....
}

Does that "exclusive" mean that only one wait item (process or thread in userspace) will be waked up?

But when I wrote some test code, I saw that the thundering herd problem still exists.

And WHY can't it be solved? Thanks!

Answer 1

In the kernel code we can see in include/linux/wait.h that __add_wait_queue_exclusive() adds the entry to the head of the list:

 __add_wait_queue_exclusive(wait_queue_head_t *q, wait_queue_t *wait)
{
        wait->flags |= WQ_FLAG_EXCLUSIVE;
        __add_wait_queue(q, wait);
}

When it comes to waking up static void __wake_up_common() in sched/wait.c does wake only the first tasks that are not exclusive and the first exclusive one. So normally only one tasks gets woken up.

Answer 2

It depends on whether you are using the same epfd. The flag WQ_FLAG_EXCLUSIVE is works only for the TASKs waiting on the same eventpoll.

If your testing code use different epfd monitor on the same socket, YES, the issue exists.